Lightcap/multi-domain-cloudflare-observability

Multi-Domain Cloudflare Web Traffic, Performance and Security Observability Dataset

This dataset contains multi-domain Cloudflare analytics exported into analysis-ready Parquet tables. It combines HTTP request aggregates, hourly traffic trends, path and referrer dimensions, country/device/browser breakdowns, DNS analytics, cache behavior, Web Vitals/RUM performance signals, redirect and ruleset metadata, and available firewall/security aggregates across 200 websites.

The collector does not use one global history window. Each zone and Cloudflare dataset is queried from the earliest retained interval exposed by Cloudflare settings.notOlderThan, clipped by the zone's Cloudflare created_on timestamp when available. The snapshot was generated at 2026-05-01T10:10:03Z and the collection end is 2026-05-01T10:00:00Z UTC.

Current observed coverage: http_daily spans 2025-05-02 to 2026-05-01; http_hourly spans 2026-01-31T13:00:00Z to 2026-05-01T09:00:00Z. Exact per-table ranges are in metadata/collection_manifest.json.

The http_daily table includes 46,108 rows from 2025 and 23,598 rows from 2026. Hugging Face's first/default dataset config is intentionally http_daily so the retained 2025 traffic is visible immediately.

Curated by Faruk Alpay. Contact: alpay@lightcap.ai.

The shape is intentionally relational: domains live in zones, redirects live in redirect_list_items, hourly and daily traffic live in http_hourly and http_daily, and high-cardinality request slices live in http_dimension_groups. That makes the data easy to query with DuckDB, Polars, Pandas, Spark, or datasets.load_dataset() without unpacking nested API responses.

Cloudflare GraphQL field availability is discovered per zone before querying. Unsupported firewall fields are not queried, so the public tables do not contain API failure rows; coverage and skipped field lists are recorded in metadata/collection_manifest.json and metadata/cloudflare_graphql_capabilities.json.

Table Overview

• http_daily: 69,706 rows
• http_hourly: 12,112 rows
• http_dimension_groups: 310,912 rows
• firewall_events: 70,322 rows
• dns_analytics: 87,482 rows
• rum_web_vitals_groups: 165 rows
• rum_performance_groups: 278 rows
• rum_pageload_groups: 303 rows
• zones: 200 rows
• accounts: 1 rows
• dns_records: 288 rows
• zone_settings: 11,200 rows
• rulesets: 783 rows
• ruleset_rules: 36,664 rows
• redirect_lists: 1 rows
• redirect_list_items: 50 rows
• web_analytics_sites: 12 rows
• web_analytics_rules: 12 rows
• metric_availability: 6 rows
• data_dictionary: 42 rows
• email_routing_rules: 201 rows
• http_hourly_browser_map: 10,259 rows
• http_hourly_content_type_map: 25,178 rows
• http_hourly_country_map: 37,120 rows
• http_hourly_http_version_map: 16,621 rows
• http_hourly_ip_class_map: 18,125 rows
• http_hourly_ssl_map: 23,821 rows
• http_hourly_status_map: 25,825 rows
• http_hourly_threat_pathing_map: 145 rows
• speed_availability: 200 rows
• speed_pages: 1 rows
• ssl_certificate_packs: 200 rows

schema/data_dictionary.json mirrors the table descriptions and important column notes. metadata/collection_manifest.json records the collection window, Cloudflare API surfaces, limits, and redaction boundary.

Example Use Cases

• Web traffic anomaly detection across multiple domains
• Hourly and daily seasonality modeling
• Bot and suspicious traffic trend analysis
• Cache hit/miss and CDN behavior analysis
• Country, ASN, device, browser and operating-system traffic breakdowns
• Core Web Vitals and real-user performance monitoring
• DNS query pattern analysis
• Security event aggregation and firewall action trend exploration
• Benchmarking tabular/time-series pipelines with real operational web data

Quickstart

from datasets import load_dataset

daily = load_dataset("Lightcap/multi-domain-cloudflare-observability", "http_daily", split="train")
hourly = load_dataset("Lightcap/multi-domain-cloudflare-observability", "http_hourly", split="train")
paths = load_dataset("Lightcap/multi-domain-cloudflare-observability", "http_dimension_groups", split="train")
zones = load_dataset("Lightcap/multi-domain-cloudflare-observability", "zones", split="train")

INSTALL httpfs;
LOAD httpfs;

SELECT
  zone_name,
  CAST(date AS DATE) AS day,
  sum(requests) AS requests,
  max(unique_visitors) AS daily_unique_visitors
FROM read_parquet('hf://datasets/Lightcap/multi-domain-cloudflare-observability/data/http_daily.parquet')
WHERE CAST(date AS DATE) < DATE '2026-01-01'
GROUP BY 1, 2
ORDER BY requests DESC;

import polars as pl

df = pl.scan_parquet("hf://datasets/Lightcap/multi-domain-cloudflare-observability/data/http_dimension_groups.parquet")

top_paths = (
    df.filter(pl.col("grouping") == "path")
      .group_by(["zone_name", "host", "path"])
      .agg(pl.sum("requests").alias("requests"))
      .sort("requests", descending=True)
      .collect()
)

print(top_paths.head())

Coverage And Limitations

• Cloudflare exposes aggregate visitor metrics here, not person-level sessions. unique_visitors is the Cloudflare uniq.uniques value for the chosen interval.
• RUM performance tables include page-load and rendering timings. They do not represent how long a visitor stayed on a page.
• Returning visitors, dwell time, bounce rate, and exit pages are not published as direct Cloudflare metrics in this export. metric_availability records those gaps explicitly instead of filling them with estimates.
• Cloudflare exposes different analytics fields depending on zone, product, plan and retention window. The dataset keeps stable schemas; unsupported fields stay null and are documented in metadata/cloudflare_graphql_capabilities.json.
• Long-tail dimension tables are bounded by the configured Cloudflare API limits. Rows with is_limit_boundary = true reached the requested API limit and should be treated as potentially truncated.

Privacy And Safety

• Full client IPs, full user-agent strings, query strings, session hashes, and hidden proxied DNS origins are excluded from the public files.
• Firewall event IPs and user agents, when available from Cloudflare, are represented as stable hashes rather than raw values.
• Proxied DNS record content is represented by a stable SHA-256 hash. Public redirect targets remain visible because they are part of the observable redirect graph.

Cloudflare references: GraphQL Analytics API, Web Analytics metrics.

Suggested Analyses

Traffic seasonality, bot-heavy domains, redirect concentration, anomaly detection, domain portfolio routing, DNS hygiene, cache behavior, country/ASN distribution, status-code drift, and launch-date inference from first observed traffic all work directly from the published tables.